Article 57 requires Member States to ensure that their competent authorities establish at least one AI regulatory sandbox at national level, with the Service Desk wording that Member States should have at least one operational by 2 August 2027, and with the consolidated Article 57 text as amended to be checked for the operative duty. Article 58 sets detailed arrangements for the establishment and functioning of sandboxes, including eligibility, selection, supervision, duration and reporting. Article 59 provides for further processing of personal data for developing certain AI systems in the public interest in a sandbox, subject to safeguards and to data protection law. A sandbox operates under a sandbox plan, under regulatory supervision, for a limited time, and ends with an exit report. It does not authorise placing on the market.
What must a competent authority establish?
Article 57(1), as amended by Regulation (EU) 2026/1744, requires Member States to ensure that their competent authorities operate at least one national AI regulatory sandbox, with the details of the duty, including any Union level sandbox provision, to be checked in the consolidated text. The sandbox is established by the competent authority, which must allocate sufficient resources, involve other relevant authorities including data protection authorities, market surveillance authorities and sector authorities, and publish information on the sandbox. The authority must also establish a single European interface or information point where relevant information on sandboxes is available, in accordance with the Article 57 provisions on information. A sandbox is therefore a public, supervised framework, not a private testing arrangement. A provider cannot establish its own sandbox, and a testing arrangement that is not established by a competent authority, that has no sandbox plan, and that has no authority supervision, is not a regulatory sandbox under the Act, however controlled the testing may be. The establishment position in each Member State, including the authority, the sandbox notice, the eligibility criteria and the application route, must be checked with the national competent authority, because sandboxes are national in establishment even where information is shared at Union level.
What are the detailed arrangements in Article 58?
Article 58 sets out detailed arrangements for, and the functioning of, AI regulatory sandboxes. The competent authority must set eligibility criteria and a selection procedure, must agree a sandbox plan with the provider, must supervise the testing, must provide guidance on the regulatory expectations, including on fundamental rights, health and safety, and must ensure that risks are identified, assessed and mitigated. The sandbox plan must set out the system, the testing, the duration, the participants, the data, the safeguards, the monitoring and the exit arrangements. The authority must also ensure that the sandbox is operated in a way that is transparent, with information published on the sandbox, the participants and the outcomes, subject to confidentiality. The Commission may adopt implementing acts to specify the detailed arrangements, and the Article 58 text, including any implementing acts, must be checked for the current position. A provider that applies to a sandbox should prepare an application that addresses the eligibility criteria, the innovation, the risk, the safeguards and the learning objectives, with the evidence for each, and should be prepared for the authority to ask questions, to require changes to the plan, and to supervise the testing closely. An application that addresses only the commercial case will not meet the criteria.
For the broader regulatory framework, see codes, standards and sandboxes overview.
How is personal data handled in a sandbox?
Article 59 provides for further processing of personal data for developing certain AI systems in the public interest in a regulatory sandbox, subject to conditions and safeguards. The further processing must be for a public interest purpose of a kind listed in the Article, such as public safety, public health, environmental protection or energy sustainability, and must meet the conditions in the Article, including that the data was lawfully collected for another purpose, that the further processing is necessary, that appropriate safeguards are in place, and that the data protection authority is involved where required. Article 59 does not create a general permission to process personal data in a sandbox, and it does not remove the need for a lawful basis, for data minimisation, for security, for data subject rights, and for a data protection impact assessment where required. The sandbox plan and the data protection assessment must be consistent, with the same data categories, purposes, retention and safeguards described in both. A sandbox application involving personal data, and in particular special category data, should involve the data protection adviser early, and should record the Article 59 conditions relied on, the safeguards, and the data protection authority involvement. A provider that assumes that sandbox supervision replaces data protection compliance will fail both assessments.
What supervision and safeguards apply during testing?
During sandbox testing, the competent authority supervises the provider, provides guidance on risk assessment and mitigation, and can require changes to the plan, suspension of testing, or termination where risks materialise or where the plan is not followed. The provider must implement the safeguards in the plan, monitor the testing, report incidents and serious incidents, and keep records that allow the authority to follow the testing. Testing in real world conditions within a sandbox must comply with the conditions in Article 60, as provided in Article 76, including the real world testing plan, the safeguards, the monitoring and the informed consent arrangements where Article 61 requires them. Participants in real world testing must be informed, must give informed consent where required, and must be able to withdraw. Vulnerable persons must be protected, and testing must not be conducted where the risks are unacceptable. The provider remains liable for damage caused, subject to the specific provision in Article 57 on administrative fines, which must be read in full and should not be overstated as a general immunity. The supervision record, including authority guidance, provider reports and any directions, should be kept with the sandbox file, so that the exit report can be based on a complete record.
What happens after sandbox testing?
After sandbox testing, the competent authority drafts an exit report, describing the testing, the outcomes, the compliance lessons and any further action, and the provider contributes the evidence and analysis for the report. The exit report is also dealt with in a separate guide in this cluster on sandbox plans and exit reports. The authority must also report annually on sandbox activity, including the number of sandboxes, the participants, the outcomes and the lessons learned, to the Commission and the Board, in accordance with Article 57. The exit report does not certify the system and does not authorise placing on the market. It provides a supervised learning record that the provider can use in its conformity assessment, its risk management and its post market planning. A provider that has completed a sandbox should update its conformity file with the exit report, the evidence produced, and the changes made to the system as a result of the sandbox learning, with the version history recorded. A provider that does not update its file will lose the value of the sandbox, because the learning will not be visible to a notified body or a market surveillance authority.
Continue with real world testing.
What does this guide not decide?
This guide does not decide whether a sandbox is available in any Member State, does not state the eligibility criteria for any sandbox, and does not reproduce Articles 57 to 59 or any implementing acts in full. Sandbox availability, criteria and application routes must be checked with the national competent authority on the date of application, and the consolidated Articles and any implementing acts must be checked for the current position. The sandbox date wording, that Member States should have at least one sandbox operational by 2 August 2027, should be kept in its Service Desk form and not upgraded without checking the consolidated Article 57 text as amended. A provider with a specific testing proposal should confirm the sandbox route, the plan and the safeguards with the competent authority before testing begins.
Frequently asked questions
Who establishes a regulatory sandbox?
The competent authority of a Member State, under Article 57, with involvement of other relevant authorities. A provider cannot establish its own sandbox. The establishment, eligibility criteria and application route in each Member State must be checked with its national competent authority.
Can personal data be processed freely in a sandbox?
No. Article 59 is narrow and conditional, for defined public interest purposes, with safeguards and data protection law applying in full. A lawful basis, minimisation, security, rights and impact assessment are still required, and the data protection authority must be involved where required.
Does a sandbox remove liability or fines?
No, not in general. The provider remains liable for damage, subject to the specific Article 57 provision on administrative fines which must be read in full. Sandbox participation does not remove liability, does not authorise placing on the market, and does not replace conformity assessment.
What is an exit report?
A report drafted by the competent authority after sandbox testing, describing the testing, outcomes, compliance lessons and further action, based on the provider evidence. It is a learning record for use in the conformity process, not a certification or a market authorisation.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗AI Act Service Desk timeline ↗